LD-Net: An Efficient Lightweight Denoising Model Based on Convolutional Neural Network

Abstract

The removal of impulse noise is a crucial pre-processing step in image processing systems. In recent years, numerous noise-removal methods have been proposed to improve denoizing performance and reconstruct noise-free images. However, removing high-density impulse noise remains a major challenge. In this paper, to address the image denoizing problem associated with high-density noise, we propose a new denoizing model, called LD-Net, which can be trained end-to-end and directly reconstructs noise-free images via a lightweight convolutional neural network. LD-Net is performed in two stages including a feature augmentation stage and a feature refinement stage. During the feature augmentation stage, the spatial size and dimension of the input image are increased by employing the deconvolutional layers for effective feature learning. During the feature refinement stage, the textural details of the image are enhanced for the reconstruction of the noise-free image by the utilization of a proposed sequence of three convolutional layers. Quantitative and qualitative evaluations performed on the SN-LABELME dataset indicate that the proposed LD-Net removes high-density impulse noise more effectively and at higher speed than other state-of-the-art denoizing methods.